1 ;*****************************************************************************
2 ;* MMX/SSE2/AVX-optimized 10-bit H.264 iDCT code
3 ;*****************************************************************************
4 ;* Copyright (C) 2005-2011 x264 project
6 ;* Authors: Daniel Kang <daniel.d.kang@gmail.com>
8 ;* This file is part of FFmpeg.
10 ;* FFmpeg is free software; you can redistribute it and/or
11 ;* modify it under the terms of the GNU Lesser General Public
12 ;* License as published by the Free Software Foundation; either
13 ;* version 2.1 of the License, or (at your option) any later version.
15 ;* FFmpeg is distributed in the hope that it will be useful,
16 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
17 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 ;* Lesser General Public License for more details.
20 ;* You should have received a copy of the GNU Lesser General Public
21 ;* License along with FFmpeg; if not, write to the Free Software
22 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 ;******************************************************************************
25 %include "libavutil/x86/x86util.asm"
30 %define pw_pixel_max pw_1023
33 ;-----------------------------------------------------------------------------
34 ; void ff_h264_idct_add_10(pixel *dst, int16_t *block, int stride)
35 ;-----------------------------------------------------------------------------
43 CLIPW %1, %4, [pw_pixel_max]
63 IDCT4_1D d,0,1,2,3,4,5
64 TRANSPOSE4x4D 0,1,2,3,4
66 IDCT4_1D d,0,1,2,3,4,5
72 STORE_DIFFx2 m0, m1, m4, m5, %1, %3
74 STORE_DIFFx2 m2, m3, m4, m5, %1, %3
78 cglobal h264_idct_add_10, 3,3
79 IDCT4_ADD_10 r0, r1, r2
90 ;-----------------------------------------------------------------------------
91 ; void ff_h264_idct_add16_10(pixel *dst, const int *block_offset,
92 ; int16_t *block, int stride,
93 ; const uint8_t nnzc[6*8])
94 ;-----------------------------------------------------------------------------
95 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
97 add4x4_idct %+ SUFFIX:
103 IDCT4_1D d,0,1,2,3,4,5
104 TRANSPOSE4x4D 0,1,2,3,4
106 IDCT4_1D d,0,1,2,3,4,5
112 STORE_DIFFx2 m0, m1, m4, m5, r5, r3
114 STORE_DIFFx2 m2, m3, m4, m5, r5, r3
121 %if HAVE_AVX_EXTERNAL
131 call add4x4_idct %+ SUFFIX
138 %macro IDCT_ADD16_10 0
139 cglobal h264_idct_add16_10, 5,6
161 %if HAVE_AVX_EXTERNAL
166 ;-----------------------------------------------------------------------------
167 ; void ff_h264_idct_dc_add_10(pixel *dst, int16_t *block, int stride)
168 ;-----------------------------------------------------------------------------
169 %macro IDCT_DC_ADD_OP_10 3
172 paddw m1, m0, [%1+0 ]
173 paddw m2, m0, [%1+%2 ]
174 paddw m3, m0, [%1+%2*2]
175 paddw m4, m0, [%1+%3 ]
197 cglobal h264_idct_dc_add_10,3,3
204 mova m6, [pw_pixel_max]
205 IDCT_DC_ADD_OP_10 r0, r2, r1
208 ;-----------------------------------------------------------------------------
209 ; void ff_h264_idct8_dc_add_10(pixel *dst, int16_t *block, int stride)
210 ;-----------------------------------------------------------------------------
211 %macro IDCT8_DC_ADD 0
212 cglobal h264_idct8_dc_add_10,3,4,7
219 mova m6, [pw_pixel_max]
220 IDCT_DC_ADD_OP_10 r0, r2, r1
222 IDCT_DC_ADD_OP_10 r0, r2, r1
228 %if HAVE_AVX_EXTERNAL
233 ;-----------------------------------------------------------------------------
234 ; void ff_h264_idct_add16intra_10(pixel *dst, const int *block_offset,
235 ; int16_t *block, int stride,
236 ; const uint8_t nnzc[6*8])
237 ;-----------------------------------------------------------------------------
240 mov r5d, [r1+(%1+0)*4]
241 call add4x4_idct %+ SUFFIX
242 mov r5d, [r1+(%1+1)*4]
244 call add4x4_idct %+ SUFFIX
249 %assign last_block 16
250 %macro ADD16_OP_INTRA 2
256 mov r5d, [r1+(%1+0)*4]
257 call idct_dc_add %+ SUFFIX
265 %macro IDCT_ADD16INTRA_10 0
266 idct_dc_add %+ SUFFIX:
277 mova m6, [pw_pixel_max]
278 IDCT_DC_ADD_OP_10 r5, r3, r6
281 cglobal h264_idct_add16intra_10,5,7,8
282 ADD16_OP_INTRA 0, 4+1*8
283 ADD16_OP_INTRA 2, 4+2*8
284 ADD16_OP_INTRA 4, 6+1*8
285 ADD16_OP_INTRA 6, 6+2*8
286 ADD16_OP_INTRA 8, 4+3*8
287 ADD16_OP_INTRA 10, 4+4*8
288 ADD16_OP_INTRA 12, 6+3*8
289 ADD16_OP_INTRA 14, 6+4*8
303 %if HAVE_AVX_EXTERNAL
308 %assign last_block 36
309 ;-----------------------------------------------------------------------------
310 ; void ff_h264_idct_add8_10(pixel **dst, const int *block_offset,
311 ; int16_t *block, int stride,
312 ; const uint8_t nnzc[6*8])
313 ;-----------------------------------------------------------------------------
315 cglobal h264_idct_add8_10,5,8,7
321 ADD16_OP_INTRA 16, 4+ 6*8
322 ADD16_OP_INTRA 18, 4+ 7*8
330 ADD16_OP_INTRA 32, 4+11*8
331 ADD16_OP_INTRA 34, 4+12*8
338 %endmacro ; IDCT_ADD8
342 %if HAVE_AVX_EXTERNAL
347 ;-----------------------------------------------------------------------------
348 ; void ff_h264_idct8_add_10(pixel *dst, int16_t *block, int stride)
349 ;-----------------------------------------------------------------------------
395 SWAP 7, 6, 4, 5, 2, 3, 1, 0 ; 70315246 -> 01234567
398 %macro IDCT8_1D_FULL 1
405 IDCT8_1D [%1], [%1+ 64*2]
408 ; %1=int16_t *block, %2=int16_t *dstblock
409 %macro IDCT8_ADD_SSE_START 2
412 TRANSPOSE4x4D 0,1,2,3,8
414 TRANSPOSE4x4D 4,5,6,7,8
418 TRANSPOSE4x4D 0,1,2,3,7
424 TRANSPOSE4x4D 4,5,6,7,3
432 ; %1=uint8_t *dst, %2=int16_t *block, %3=int stride
433 %macro IDCT8_ADD_SSE_END 3
439 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
441 STORE_DIFFx2 m2, m3, m6, m7, %1, %3
445 STORE_DIFFx2 m4, m5, m6, m7, %1, %3
447 STORE_DIFFx2 m0, m1, m6, m7, %1, %3
451 cglobal h264_idct8_add_10, 3,4,16
453 %assign pad 16-gprsize-(stack_offset&15)
455 call h264_idct8_add1_10 %+ SUFFIX
461 ; TODO: does not need to use stack
462 h264_idct8_add1_10 %+ SUFFIX:
463 %assign pad 256+16-gprsize
468 IDCT8_ADD_SSE_START r1, rsp
475 IDCT8_ADD_SSE_START r1+16, rsp+128
476 PERMUTE 1,9, 2,10, 3,11, 5,1, 6,2, 7,3, 9,13, 10,14, 11,15, 13,5, 14,6, 15,7
477 IDCT8_1D [rsp], [rsp+128]
486 IDCT8_1D [rsp+16], [rsp+144]
508 CLIPW m8, m0, [pw_pixel_max]
510 mova m8, [pw_pixel_max]
511 STORE_DIFF16 m9, m1, m0, m8, r0+r2
513 STORE_DIFF16 m10, m2, m0, m8, r0
514 STORE_DIFF16 m11, m3, m0, m8, r0+r2
516 STORE_DIFF16 m12, m4, m0, m8, r0
517 STORE_DIFF16 m13, m5, m0, m8, r0+r2
519 STORE_DIFF16 m14, m6, m0, m8, r0
520 STORE_DIFF16 m15, m7, m0, m8, r0+r2
522 IDCT8_ADD_SSE_START r1, rsp
523 IDCT8_ADD_SSE_START r1+16, rsp+128
525 IDCT8_ADD_SSE_END r0, rsp, r2
526 IDCT8_ADD_SSE_END r3, rsp+16, r2
551 %if HAVE_AVX_EXTERNAL
556 ;-----------------------------------------------------------------------------
557 ; void ff_h264_idct8_add4_10(pixel **dst, const int *block_offset,
558 ; int16_t *block, int stride,
559 ; const uint8_t nnzc[6*8])
560 ;-----------------------------------------------------------------------------
561 ;;;;;;; NO FATE SAMPLES TRIGGER THIS
562 %macro IDCT8_ADD4_OP 2
567 call h264_idct8_add1_10 %+ SUFFIX
575 cglobal h264_idct8_add4_10, 0,7,16
576 %assign pad 16-gprsize-(stack_offset&15)
583 IDCT8_ADD4_OP 0, 4+1*8
584 IDCT8_ADD4_OP 4, 6+1*8
585 IDCT8_ADD4_OP 8, 4+3*8
586 IDCT8_ADD4_OP 12, 6+3*8
589 %endmacro ; IDCT8_ADD4
593 %if HAVE_AVX_EXTERNAL